Minglei Li

you can collect robot training data without robots now HiFi-UMI records humans performing tasks with a handheld gripper, then uses that data to post-train policies that deploy directly on real robots Across 3 model families, it matched robot teleoperation within 3.1 percentage points Best result: 85% success on precision insertion. it's just a matter of collecting accurate human demonstrations now.

Stop putting a person in a teleoperation rig to train a robot. A new system called HiFi-UMI captures manipulation data without a robot in the loop, using a handheld gripper rig carried through the task. It solves the accuracy problem that made this shortcut unreliable until now. > 3mm end effector tracking accuracy > Under 40 microseconds of camera sync > Two wide angle cameras per hand A policy trained only on this handheld data matched real teleoperation, with success gaps of about 3 percentage points or less across three model families. The strongest version reached 85 percent on a precision task, despite zero demonstrations from that scene. Pretraining on 4,000 hours of the same data cut action error by 41 percent on unseen tasks. The team released 2,000 hours of it publicly under an open license. What breaks first once teleoperation stops being the bottleneck?